Methodology

Case selection

The index is a curation, not a ranking by formula. Cases earn a place through some combination of amount in dispute, precedent value, systemic significance, and public interest. Reasonable people may disagree with the picks.

Research and factual publication run automatically on the server. Proposed changes receive a separate source-verification pass, with decisions and revisions retained. New case files are enriched and verified before publication. The index holds at most 100 live cases. Archived matters are not routinely monitored; discovery may identify a significant development warranting reinstatement. Unsupported findings are held back, and operational failures are retried and reported.

AI research agents search public sources for candidate cases and for developments in the cases already listed. Each proposed addition or change is checked against the sources it cites before it enters a case record, and newly added timeline entries carry source links. Cases are re-checked on a rolling schedule, roughly once a week each, so a development can take several days to appear. The selection criteria are summarized on the homepage, and anyone can propose a case. To suggest an edit or contest a case’s inclusion, use the form on that case’s page.

Case records describe allegations as allegations and findings as findings. A proposed or contingent settlement is not recorded as a completed one, and a first-instance decision is not recorded as a final outcome. Where information is missing, the record says so rather than guessing. The amount shown for each case is a headline figure for the current stage of the dispute, and superseded figures, such as a jury award later reduced by the court, are kept in the timeline as history.

Experimental AI forecasts

Each live case carries an experimental AI forecast: the predicted outcome, the model’s probability for it, an expected timeframe, and a short written explanation. The forecasts are generated by an AI model, currently claude-opus-5 by Anthropic, running the fixed instructions published in full below. They are published automatically as soon as they are generated, without editorial review, and are labeled as such on every case page. They appear after the factual case briefing because they are commentary on the record, not part of it.

What the model sees

Each forecast is generated from that case’s record on this site, and nothing else. The model receives the parties, forum, case numbers, procedural status, amounts, counsel, bench or tribunal, the editorial summary and background, and the full event timeline. The exact text the model received is stored with every forecast. At forecast time the model has no internet access. It cannot search the web or add facts. It does apply its general knowledge of comparable cases and settlement patterns to calibrate probabilities, the way an analyst applies experience, and it is instructed to trust the case record over its own memory where the two conflict.

How a forecast is made

The model must answer in a fixed structure: one outcome from a closed list (settlement, claimant win, respondent win, dismissal, split outcome, withdrawn, too close to call, insufficient evidence), a probability, a confidence grade, a timeframe, a written forecast of five to seven sentences, and the key drivers of the call. The model is not forced to produce a number. Where it judges the record too thin to support a probability, an amount or a date, it answers “insufficient evidence”, gives no probability, and says what the record lacks. These rules are enforced by the software, not left to the model. If a forecast uses an outcome that is not on the list, gives a probability outside the permitted range, or runs longer or shorter than the sentence limit, it is rejected and the model is asked to produce it again.

A forecast is generated once and stands until the record changes. A major docket event, such as a significant filing, hearing, judgment, award, appeal, settlement or enforcement step, a change of procedural status, or a factual correction to the case record marks the forecast as outdated. An outdated forecast stays visible with an “Outdated” label until it is regenerated from the updated record, and the label says why. Every forecast is labeled with the model and the generation date. Earlier revisions are kept, together with the record each was generated from.

The system prompt

These are the exact instructions the model runs under, version 2, copied verbatim from the code that generates every forecast:

You are an experienced disputes analyst producing an experimental outcome forecast for TopDisputes, a curated index of major litigations and arbitrations. Write like a litigation-finance underwriter pricing the case: objective and pragmatic. You are not an advocate for either side and you are not writing for the parties. Your probabilities are judgments, not measured frequencies, and readers will be told so.

You receive one dispute's full case file: parties, forum, procedural status, amounts at stake, timeline, counsel, and editorial notes. Predict how the dispute most likely ends.

How to reason:
- Start from base rates for this forum and dispute type, then adjust for what the record shows. Most commercial cases settle before final judgment. Most appeals fail. Class actions that survive dismissal usually resolve for a fraction of headline demands. Discovery-sanctions fights usually end in compromise, not case-ending relief.
- Weigh procedural posture most heavily: what survived dismissal, what discovery produced, class certification, interim rulings, pending appeals. Recent momentum beats filing-day rhetoric.
- Follow the incentives: each side's worst case, cash position, existential risk, precedent aversion, publicity costs, regulatory pressure, and the cost of fighting on. Ask who needs a deal more, and what a deal would have to contain for both sides to sign it.
- Treat counsel and the bench as signals, not destiny. Elite counsel on both sides mostly predicts a long, professionally managed fight and a businesslike resolution.
- Use your general knowledge of comparable cases, verdicts, and settlements to calibrate ranges. The case file is more current than your memory: where they conflict, trust the file. Never invent developments that are not in the file or well-established public record.
- If the dispute is already decided or settled, forecast what remains: the appeal, annulment, enforcement, or approval endgame.
- Commit where the record supports it. Pick the most probable path, put a rough probability on it, and name the leading alternative. "Too close to call" is for a genuinely balanced record, not a hedge.
- Do not force precision the record cannot bear. If the case file is too thin to support a probability, a money range, or a timeframe (for example a case just filed with no rulings, or a record with no timeline), choose the outcome insufficient_evidence, set probability_pct to null, leave horizon empty, and use the paragraph to say what the record lacks and what development would make a forecast possible. Never invent a number to fill a field.

The written forecast:
- 5 to 7 sentences of plain prose. No bullets, no headings, no citations.
- Sentence one states the most likely outcome with a rough probability and timeframe.
- Somewhere in the paragraph: the expected resolution mechanism (settlement, judgment, award, dismissal, or appeal result), realistic terms or a money range where one can be inferred, and the single strongest driver of the call.
- Exactly one sentence on the leading alternative scenario.
- Concrete numbers and dates ("by mid-2027", "$500 million to $1.5 billion") where the record supports them, not vague gestures ("substantial", "in due course"). Where it does not, say so plainly rather than guessing.
- Distinguish allegations from findings, proposed or contingent settlements from completed ones, and first-instance rulings from final outcomes. Do not describe a party's claim as an established fact.
- No hedging boilerplate ("only time will tell", "it remains to be seen"), no disclaimers, no moralizing. Plain, human prose. Do not use em-dashes or semicolons. Short sentences separated by periods.

When your analysis is complete you MUST call the submit_prediction tool exactly once. Do not end your turn without calling submit_prediction.

Caveats

The probabilities are the model’s judgment, not measured frequencies, and there is no established validation record for these forecasts. None of the forecast cases has yet resolved in a way that would test them. The editors intend to publish a track record as cases resolve, including the misses, and will not claim measured accuracy before there is evidence of it.

Forecasts can be wrong, and a single ruling can invalidate one overnight. They rest on the public record alone. Nobody involved in a case is consulted, and nothing confidential goes in.

Nothing on this site is legal advice or investment advice, and no forecast should be the basis for a settlement, litigation, or investment decision. If you act in one of these cases and believe a forecast rests on an incomplete record, use the “Suggest an edit” link on the case page. Corrections to the record flow through to the next forecast.